C - MPI - Send/Receive 子数组到数组

C - MPI - Send/Receive Subarrays to Array

假设我们有一个包含 master_array 个 6*6 单元格的主 MPI 进程:

  Master
-----------
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0

并且我们有 4 个工作 MPI 进程,其中 worker_array 个 3*3 个单元格。

Worker 1 | Worker 2 | Worker 3 | Worker 4 |
-------  | -------  | -------  | -------  |
 1 1 1   |  2 2 2   |  3 3 3   | 4 4 4    |
 1 1 1   |  2 2 2   |  3 3 3   | 4 4 4    |
 1 1 1   |  2 2 2   |  3 3 3   | 4 4 4    |

现在,我想像这样将工作数组发送到主数组:

  Master
-----------
1 1 1 2 2 2
1 1 1 2 2 2
1 1 1 2 2 2
3 3 3 4 4 4 
3 3 3 4 4 4 
3 3 3 4 4 4 

如何使用 some-kind-of-MPI-send/receive、MPI_datatypes 或 MPI_vectors 或 MPI_subarrays 结束此操作?

将不胜感激提供详细且有效代码的答案。

这是一个同时使用点对点和集体的工作代码(集体版本在下面被注释掉但工作正常)。您需要定义一个向量类型来对应主控端接收端的非连续数据。要使用集体收集,您需要弄乱此向量的大小以确保收集将所有部分放在正确的位置并且您需要使用 gatherv 版本。

数组索引很容易弄乱,所以为了一般性,我在 6x12 矩阵上使用了 2x3 进程数组,因此故意不正方形。

为乱七八糟的缩进道歉 - 我似乎有 tab/space 个问题,我真的应该在以后解决!

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

#define M 6
#define N 12

#define MP 2
#define NP 3

#define MLOCAL (M/MP)
#define NLOCAL (N/NP)

#define TAG 0

int main(void)
{
  int master[M][N];
  int local[MLOCAL][NLOCAL];

  MPI_Comm comm = MPI_COMM_WORLD;
  int rank, size, src;
  int i, j;
  int istart, jstart;
  int displs[MP*NP], counts[MP*NP];

  MPI_Status status;
  MPI_Request request;
  MPI_Datatype block, blockresized;

  MPI_Init(NULL, NULL);

  MPI_Comm_size(comm, &size);
  MPI_Comm_rank(comm, &rank);

  if (size != MP*NP)
    {
      if (rank == 0) printf("Size %d not equal to MP*NP = %d\n", size, MP*NP);
      MPI_Finalize();
      return 1;
    }

  for (i=0; i < M; i++)
    {
      for (j=0; j < N; j++)
    {
      master[i][j] = rank;
    }
    }

  for (i=0; i < MLOCAL; i++)
    {
      for (j=0; j < NLOCAL; j++)
    {
      local[i][j] = rank+1;
    }
    }

  // Define vector type appropriate for subsections of master array

  MPI_Type_vector(MLOCAL, NLOCAL, N, MPI_INT, &block);
  MPI_Type_commit(&block);

  // Non-blocking send to avoid deadlock with rank 0 sending to itself

  MPI_Isend(local, MLOCAL*NLOCAL, MPI_INTEGER, 0, TAG, comm, &request);

  // Receive from all the workers

  if (rank == 0)
    {
      for (src=0; src < size; src++)
    {
      // Find out where this block should go

      istart = (src/NP) * MLOCAL;
      jstart = (src%NP) * NLOCAL;

      // receive a single block

      MPI_Recv(&master[istart][jstart], 1, block, src, TAG, comm, &status);
    }
    }

  // Wait for send to complete

  MPI_Wait(&request, &status);

  /* comment out collective

  // Using collectives -- currently commented out!

  MPI_Type_create_resized(block, 0, sizeof(int), &blockresized);
  MPI_Type_commit(&blockresized);

  // Work out displacements in master in counts of integers

  for (src=0; src < size; src++)
    {
      istart = (src/NP) * MLOCAL;
      jstart = (src%NP) * NLOCAL;

      displs[src] = istart*N + jstart;
      counts[src] = 1;
    }

  // Call collective

  MPI_Gatherv(local, MLOCAL*NLOCAL, MPI_INT,
          master, counts, displs, blockresized,
          0, comm);

  */

  // Print out

  if (rank == 0)
    {
      for (i=0; i < M; i++)
    {
      for (j=0; j < N; j++)
        {
          printf("%d ", master[i][j]);
        }
      printf("\n");
    }
    }

  MPI_Finalize();
}

它似乎在 6 个进程上工作正常:

mpiexec -n 6 ./arraygather
1 1 1 1 2 2 2 2 3 3 3 3 
1 1 1 1 2 2 2 2 3 3 3 3 
1 1 1 1 2 2 2 2 3 3 3 3 
4 4 4 4 5 5 5 5 6 6 6 6 
4 4 4 4 5 5 5 5 6 6 6 6 
4 4 4 4 5 5 5 5 6 6 6 6 

这应该适用于矩阵准确分解到过程网格的任何情况。如果进程的子矩阵大小不完全相同,情况会更复杂一些。